Optimize BpNN using new breeder genetic algorithm

In this paper, the ability of genetic algorithms in designing artificial neural network (ANN) is investigated. The multi-layer network (MLN) is taken into account as the ANN structure to be optimized. The idea presented here is to use the genetic algorithms to yield contemporaneously the optimization of: (1) the design of NN architecture in terms of number of hidden layers and of number of neurons in each layer; and (2) the choice of the best parameters (learning rate, momentum term, activation functions, and order of training patterns) for the effective solution of the actual problem to be faced. The back-propagation (BP) algorithm, which is one of the best-known training methods for ANNs, is used. To verify the efficiency of the current scheme, a new version of the breeder genetic algorithm (NBGA) is proposed and used for the automatic synthesis of NN. Finally, several problems of the experiment were taken and the results show that the back-propagation neural network (BpNN) classifier improved the current scheme has higher accuracy of classification and greater gradient of convergence than other classifiers, which have been proposed in the literature